enhance: wait for collection readiness via notifications

[sunby]

Automatic DQL readiness polls ShowLoadCollections every 10 ms while a collection loads, with separate loops for concurrent requests. Replace those loops with a WaitCollectionReady RPC signaled by committed load-config changes and shard-state updates. Share load submission and readiness waiting per collection on each Proxy, including requests that already observe Loading, while retaining the direct DQL loading path from #15.

The RPC checks every expected vchannel/replica pair. Subscribing before reading state prevents missed notifications; release invalidates existing subscriptions, and cancellation, load timeout and runtime shutdown end the RPC. Notification state is retained only for active waiters. Initial per-request readiness checks and explicit load-progress APIs remain available.

Related: #16 and #25. Issue tracking is disabled for this fork.
Design: Collection Readiness.

Upgrade MixCoord before Proxy: older coordinators do not implement the new RPC, and there is no polling fallback. The synchronous warmup policy in draft #17 remains separate from this PR.

Validation

Validated on base 7050f6f80e with -tags dynamic,test -gcflags="all=-N -l" and the installed native libraries:

• 35 selected QueryCoord, Proxy, notification, load-config and shard-registry tests passed with -race -count=3, including real StreamingNode Up callbacks, gRPC deadlines/status, release, cancellation during a blocked catalog write, shared waits and automatic-load/requery queue regressions.
• All 153 tests in the notification, load-config and shard-registry packages passed.
• MixCoord service/client, coordinator and DataCoord packages passed compile-only checks.
• Lint restricted to this diff: 0 issues; git diff --check passed.
• run_clang_format.sh passed with clang-format 15.0.7.
• make lint-fix ran gofumpt/gci, then failed on the unchanged internal/metastore/kv/querycoord/kv_catalog_test.go:372: undefined: mocks; pkg/client lint stages were not reached. Unrelated formatter changes were excluded.

Production CPU/allocation savings and mixed-version rollout have not been measured.